Repository navigation
Conversation
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes 7714.
As noted in the issue, we get flaky failures in spectral clustering tests, because Lanczos sometimes ends up having to find eigenvalues in a very tricky matrix and the cuda 12.x cusolverDnXsyevd occasionally fails. This approach falls back from the divide-and-conquer approach (the d on the end of the function name) to a Jacobi approach. It keeps things in float32.
Tried 40000 hypothesis test variations (5 of them actually fell back to jacobi) plus 176 variations that are known to fail with the original algorithm. All pass now. Note that they all already work on CUDA 13 cusolver - this is a CUDA 12 only issue.
This change potentially affects spectral embedding and umap spectral initialization, but it seems... fine? I don't see why one of those callers would care that it in rare cases used a different eigensolver algorithm.
Previous callers of eigJacobi do not check convergence. To minimize disruption, I made the new convergence check-and-throw optional.